Few-Shot and Zero-Shot Multi-Label Learning for Structured Label Spaces.

Area of Science:

• Natural Language Processing
• Machine Learning
• Computational Linguistics

Background:

• Large multi-label datasets present challenges with label imbalance, including frequent, few-shot, and zero-shot labels.
• Multi-label few- and zero-shot learning for text classification, particularly in large label spaces, remains underexplored.

Purpose of the Study:

• To conduct a fine-grained evaluation of existing state-of-the-art methods on infrequent labels in multi-label text classification.
• To develop novel few-shot and zero-shot learning methods tailored for multi-label text classification, leveraging known label space structures.

Main Methods:

• Evaluation of current methods on infrequent labels within large-scale datasets.
• Development and application of new few-shot and zero-shot learning algorithms for structured label spaces.
• Testing on publicly available medical text datasets: MIMIC II and MIMIC III.

Main Results:

• Significant improvements in R@10 for few-shot labels: 6.2% on MIMIC II and 4.8% on MIMIC III.
• Substantial gains in R@10 for zero-shot labels: 17.3% on MIMIC II and 19% on MIMIC III.
• Demonstrated effectiveness of developed methods over prior approaches for handling infrequent labels.

Conclusions:

• The proposed methods effectively enhance multi-label text classification performance for both few-shot and zero-shot labels.
• Leveraging label structure is crucial for improving prediction accuracy in scenarios with data scarcity.
• The findings highlight the potential for advancing machine learning applications in specialized domains like medical informatics.